New software unveiled last month suggests hospitals could make patient-specific predictions of antibiotic susceptibility.
Keith W. Hamilton, MD
New research suggests that in the near future antibiograms can transform from a general guideline to a patient-specific tool for predicting antibiotic susceptibility.
Investigators from the University of Pennsylvania and the health IT company ILUM Health Solutions unveiled their findings last month at the SHEA 2018 Spring Conference, in Portland, Oregon. The team stated that their model can leverage patient health information, along with hospital-specific susceptibility rates, to improve the accuracy of antibiograms.
Choosing the right antibiotic can be critical for a number of reasons. First, it can help improve patient outcomes and decrease the amount of time it takes for a patient to improve. For example, the investigators noted that appropriate empiric antibiotic treatment in septic patients led to a significant reduction in all-cause mortality.
Second, more targeted antibiotic use could minimize adverse effects of antibiotics, including the spread of antibiotic resistance.
Antibiograms have long been in use at the hospital level. These antibiograms are based on periodic analysis of antimicrobial susceptibilities of local bacterial isolates from a hospital’s own laboratory.
In this new study, the investigators went a step further, combining that data with information from patient electronic health records (EHRs) in order to refine recommendations down to a patient-by-patient prediction.
“The predictive models have been operationalized into a software platform that automatically extracts the relevant data from the electronic health record,” said Keith W. Hamilton, MD, associate healthcare epidemiologist, and director of antimicrobial stewardship at the Hospital of the University of Pennsylvania. “These data elements include prior antibiotic exposure, prior microbiology, history of resistant pathogens, diagnoses, age, and location in the hospital.”
Dr. Hamilton told Contagion® that the system is able to interface with nearly every type of EHR, meaning it could be deployed in a wide array of health care facilities.
The model was developed using a training set that incorporated clinical and microbiological data spanning from May 2008 through August 2016. Institution-specific antibiotic susceptibilities of greater than 98% and less than 2% were excluded from the data, as were organisms where the sample size was less than 10.
The evaluation set of data was collected over the last 5 months of 2016 and was made up of patients who were not included in the training data set.
The investigators compared hospital-level antibiogram susceptibility data to patient-specific data, coming up with a Brier score for each. A Brier score is a measure of the accuracy of probability-based predictions. The lower the Brier score, the higher the chance of a prediction being accurate.
The data set began with more than 11,000 positive specimens from more than 7,500 patients. That translated into 248 organism/antibiotic pairs, a number that was winnowed down by excluding outliers, resulting in 116 evaluable organism/antibiotic pairs and 111,732 patient-specific antibiogram predictions.
Patient-specific antibiograms beat or tied the predictions based on institution-level antibiograms 97% of the time. Of those, patient-specific antibiograms outperformed institution-level antibiograms about half of the time (47.6%) and showed no statistical difference the rest of the time (49.4%). The institution-level antibiogram served as a better predictive tool of antibiotic susceptibility in only 5 organism/antibiotic pairs.
Dr. Hamilton sees these patient-specific predictions as part of a larger effort to bring Big Data precision to the clinic.
“The goal of developing these sorts of software innovations is to provide medical providers with all the information they need to make the most informed decision possible for the patient in front of them,” he said. The model takes a wealth of data and converts it to an easily readable and usable format for physicians and other practitioners, he said. However, he does not see these models as a replacement for professional judgment.
“There is no substitute for a trained medical provider in weighing the risks and benefits of appropriate prescriptions, sending the appropriate diagnostic tests, and coming up with the correct diagnosis,” he said. “However, we see this software as augmenting the performance of medical providers. It can give them the data they need when they need it the most.”
One limitation of the study is that it was based on a single health system. Dr. Hamilton stated that a larger-scale study would be needed in order to understand the full impacts of large-scale deployment of patient-specific antibiograms, or on the use of antibiograms at other types of facilities.
When Contagion® asked Dr. Hamilton what surprised him the most about the findings, he noted the quick pace of data-based patient-specific technology.
“I think we all have been surprised with the improvement that predictive analytics can have over existing data sources, and this is just the beginning,” he said. “As the medical community develops new rapid diagnostic tests and can extract more data elements, these models likely can be improved to deliver even better predictions with a quicker turnaround time to make sure patients get the right antibiotic at the right time.”
Although patient-specific data could go a long way toward precision prescribing, another recent study on antibiograms points out the value of regional-level data, in addition to patient data and institutional data.
An April 23, investigators on a study published in the journal Infection Control and Hospital Epidemiology analyzed the use of antibiograms at 20 small community hospitals in eastern North Carolina and south-central Virginia.
They looked at the problem of small data sets at local hospitals, noting that small community hospitals, “frequently cannot develop an accurate antibiogram due to a paucity of local data.”
The results of their analysis showed that 69% of local hospital susceptibility rates fell within 1 standard deviation of the mean regional susceptibility rate, and 97% fell within 2 standard deviations.
The study’s authors, from Duke University and the Department of Veterans Affairs, concluded that, “A regional antibiogram is likely to provide clinically useful information to community hospitals for low-prevalence pathogens.”
Although that study advocates broader views of data, it seems to ultimately support Dr. Hamilton’s conclusion that when it comes to predicting susceptibility, the more data, the better.
Getting to the right antibiotic first, and quickly, could also help with the growing problem of antibiotic resistance. The US Centers for Disease Control and Prevention (CDC) says more than 2 million illnesses and 23,000 deaths have been caused as a result of antibiotic resistance.
“Determining appropriate empiric antimicrobial therapy for patients with infections can be challenging and depends on balancing the risks and benefits of using broad-spectrum antimicrobials versus more narrow-spectrum agents,” shared Dr. Hamilton. “Use of unnecessarily broad-spectrum antimicrobials puts patients at risk for complications, including infection and colonization with resistant organisms such Clostridium difficile, as well as drug toxicity, while the use of too-narrow-spectrum agents can result in morbidity and mortality from a delay in effective therapy.”
According to the CDC, about a quarter of a million C. difficile infections and 14,000 deaths have occurred as a result of antibiotic resistance.
Looking forward, Dr. Hamilton and colleagues say they plan to look at other clinical factors that affect predictive resistance and work to develop new ways to overcome the challenge of small sample sizes for some organisms. The team also wants to understand the broader clinical impacts of early detection of predictive resistance.
ILUM Health Solutions is a subsidiary of Merck & Co. Inc.’s Health Services and Solutions. Hamilton reported no financial disclosures. His co-authors are consultants of ILUM or employees of Merck units.